Lumos3D: A Single-Forward Framework for Low-Light 3D Scene Restoration

TL;DR

Lumos3D tackles 3D scene restoration under low light by removing the need for precomputed poses or per-scene optimization. It employs a single-forward, geometry-aware framework built on VGGT/AnySplat, with a cross-illumination distillation scheme that uses a normal-light teacher to guide a low-light student, and a Lumos loss that enforces photometric and geometric consistency. The approach reconstructs a 3D Gaussian representation via differentiable voxelization and optimizes a composite objective that includes $L_{total} = L_{rec} + \\omega_{distill} L_{distill} + \\omega_{lumos} L_{lumos}$, with multi-level losses spanning content, image, and voxel domains. Empirical results on synthetic and real datasets show high-fidelity illumination restoration and accurate geometry, with strong generalization to unseen scenes and the ability to extend to over-exposure restoration, enabling practical, real-time 3D relighting in challenging illumination conditions.

Abstract

Restoring 3D scenes captured under low-light con- ditions remains a fundamental yet challenging problem. Most existing approaches depend on precomputed camera poses and scene-specific optimization, which greatly restricts their scala- bility to dynamic real-world environments. To overcome these limitations, we introduce Lumos3D, a generalizable pose-free framework for 3D low-light scene restoration. Trained once on a single dataset, Lumos3D performs inference in a purely feed- forward manner, directly restoring illumination and structure from unposed, low-light multi-view images without any per- scene training or optimization. Built upon a geometry-grounded backbone, Lumos3D reconstructs a normal-light 3D Gaussian representation that restores illumination while faithfully pre- serving structural details. During training, a cross-illumination distillation scheme is employed, where the teacher network is distilled on normal-light ground truth to transfer accurate geometric information, such as depth, to the student model. A dedicated Lumos loss is further introduced to promote photomet- ric consistency within the reconstructed 3D space. Experiments on real-world datasets demonstrate that Lumos3D achieves high- fidelity low-light 3D scene restoration with accurate geometry and strong generalization to unseen cases. Furthermore, the framework naturally extends to handle over-exposure correction, highlighting its versatility for diverse lighting restoration tasks.

Figures (3)

• Figure 1: Architecture overview. Given multi-view low-light context inputs, Lumos3D instantly predicts 3D Gaussian representations with restored light conditions and renders corresponding RGB image and depth maps, without scene-specific training OR optimization. The two key components are the cross-illumination distillation loss $\lambda_{distill}$ and the proposed $\lambda_{lumos}$, as discussed in \ref{['sec:distill_loss']} and \ref{['sec:lumos_loss']} respectively. For simplicity, we omit the baseline loss $\lambda_{rec}$jiang2025anysplat.
• Figure 2: Qualitative comparison of different distillation schemes. Each visualization corresponds to the same scene, with depth on the left and the corresponding RGB image on the right. In the depth maps, blue denotes distant regions and red denotes nearby ones. Distillation on low-light images suffers from illumination ambiguity, whereas distillation on normal-light images yields more accurate and geometrically cleaner depth and relighting results.
• Figure 3: Qualitative comparison of different 3D low-light and over-exposure restoration schemes on the chair and sofa scenes in the LOM dataset.